Apple plans to use Amazon’s custom chips for its AI training.
Currently, Apple uses Amazon’s custom AI chips for services such as search and is evaluating whether it can leverage the latest AI chip from Amazon to pre-train its models, including Apple Intelligence.
At the AWS Reinvent annual conference on Tuesday, Apple confirmed its use of Amazon’s custom chips. Benoit Dupin, Apple’s Senior Director of Machine Learning and AI, spoke about how Apple utilizes Amazon’s cloud services. Dupin said, “We have a strong relationship, and the infrastructure is both reliable and capable of serving our customers worldwide.”
Apple’s presence at the Amazon conference and its use of Amazon chips highlighted that Cupertino is also relying on Amazon’s cloud services. Dupin mentioned that Apple has been using AWS for over a decade for services such as Siri, Apple Maps, and Apple Music. For instance, Apple has used Amazon’s Inferentia and Graviton chips for search services, ultimately achieving a 40% performance boost in its search functions.
Dupin also mentioned that Apple will use Amazon’s Trainium2 chip to pre-train its custom models. Thus, Amazon’s chips not only provide a cost-effective option for AI model inference compared to Intel and AMD x86 processors but can also be used for developing new AI models. Amazon announced on Tuesday that the Trainium2 chip is now available for public rental.
Dupin said, “In the early stages of evaluating Trainium2, we expect up to a 50% improvement in pre-training performance.” AWS CEO, Matt Garman, told CNBC that Apple is one of the early users and beta testers of Trainium chips.
Earlier this year, Apple stated in a research paper that it uses Google’s Cloud TPU chips to train its iPhone AI service, Apple Intelligence. This comes at a time when most AI training is performed on expensive Nvidia GPUs.
Cloud service providers and startups are competing to develop alternatives to reduce costs and exploring various methods to achieve more efficient processing. Apple’s use of custom chips could demonstrate to other companies that AI training methods beyond Nvidia’s solutions can also be viable.